Official: Tesla HW3 vehicles are reaching the peak of their capabilities with FSD V12.5

Tesla HW3 vehicles are reaching the peak of their capabilities
8 photos

Photo: @greentheonly via X

Tesla’s Autopilot software lead Ashok Elluswamy explained what caused the delay in rolling out FSD V12.5 to Hardware-3 (AI3) vehicles. Far from easing owner concerns, Elluswamy confirmed that the AI3 hardware is in its final moments. While Tesla can still force FSD to run on it, it may not be able to do the same with future iterations.

It looks like Tesla has abandoned Hardware-3 vehicles when it comes to FSD development. The first hardware suite capable of running Tesla Fully autonomous driving software, HW3 (or AI3, as Tesla now calls it), is present in about 80% of all Tesla vehicles currently on the road. Meanwhile, the most recent hardware iteration (HW4/AI4) has a much smaller install base, at just 12% according to TeslaFi statistics, which may or may not be representative of the entire Tesla fleet.

Despite this discrepancy, Tesla is now prioritizing FSD development for AI4 hardware. Elon Musk reassured owners that Hardware 3 had not reached its full potential, but they still had to wait more than three weeks for Tesla to release the first FSD V12.5 build for their vehicles. This is because Tesla has gone from AI3-focused development, with AI4 hardware running AI3 in emulation mode, to the exact opposite. Now, Tesla is developing FSD specifically for the AI4 computer and is looking to get it to work in AI3 vehicles.

As Elon Musk explained last month, this requires a significant amount of effort, which has led to a significant delay for AI3 vehicles. On Wednesday, Tesla finally released the 2024.26.15 software update, which includes the first FSD V12.5 build (V12.5.1.4) for AI3 vehicles. However, this is not the full V12.5 that AI4 vehicles received a few weeks ago. As Tesla’s Autopilot software chief Ashok Elluswamy explained in an X post, Tesla has been using a smaller model for AI3 vehicles.

As we learned earlier this month, Tesla is significantly increasing its parameter count, which means the in-vehicle inference computer needs to analyze many more aspects of each driving scenario. Given the computational power required, current AI3 computers would run into bottlenecks. To run the latest versions of FSD, Tesla has been forced to use a smaller software model, although Elluswamy still says AI3 should provide “similar performance” to AI4 thanks to this trick.

However, this is not a practical solution, which is why Tesla is working on implementing additional AI3 kernels in the compiler to emulate the same operations natively supported on AI4 hardware. In combination with other model compression techniques, Elluswamy thinks this would allow Tesla to “in the future we will use the same 12.5 model on both hardware platforms.”

While this works for FSD V12.5, there’s almost no guarantee it will work for future FSD iterations. With each new build, Tesla’s software and self-driving models tend to become more complex. We saw this with V12.5, which brought significant changes over V12.4. This suggests that V12.6 could bring new headaches to AI3 owners. At some point, Tesla won’t be able to backport new features to older hardware. With AI5 coming next year, Tesla will be exploring new capabilities that will make older hardware obsolete.

Leave a Reply

Your email address will not be published. Required fields are marked *